The new GPUs are meant as an alternative to having developers divide a supercomputing application among multiple CPUs, commonly made by Intel or Advanced Micro Devices.
Nvidia plans to start releasing next month a new class of processors that boost the performance of supercomputing environments by offloading trillions of arithmetic calculations per second.
Called Tesla, the new line of graphics processing units can be used in PCs, workstations or servers. Depending on the configuration, the GPUs can handle as much as 8 teraflops, or 8 trillion calculations a second.
"You don't have to divide your application into pieces," Andy Keane, general manager of GPU computing at Nvidia, told InformationWeek Thursday. "You take that small heavily computational piece and move it to the GPU, and every year we'll offer you a faster processor."
The Tesla brand is divided into three product lines. The first is the Deskside Supercomputer that attaches to a workstation or PC via a PCI Express cable. The appliance can contain up to 16 GPUs to reach a processing level of 8 teraflops.
The purpose of the new product is to turn a workstation or desktop into a mini-supercomputer, so researchers can do more in the lab, rather than have to wait or pay for time on a larger supercomputer. "They can pretty much build a supercomputer on their desktop, where they can solve a huge number of problems," Keane said.
Another Tesla product is the Computing Server, which would connect to a 1U server via a PCI Express cable. Each Nvidia server, which is the same size as a 1U server, would hold as many as eight GPUs for delivering 4 teraflops of processing power.
Finally, Nvidia is offering the Tesla Computing Processor, which is a dedicated board with one GPU. While the GPU is available on a PCI Express card for a PC, the product is more geared toward embedding in other systems. Nvidia can deliver the board in different form factors for use in medical equipment, for example.
Nvidia offers software developments tools, called Cuda, for dividing an application to run on the vendors' GPUs. The tools include a C-compiler for the GPU, debugger/profiler, dedicated driver, and standard libraries.
The Tesla Computing Server is expected to cost $12,000 with four GPUs, and the Deskside Supercomputer $7,500 with two GPUs. The Computing Processor is expected to sell for $1,499 for one GPU. The desktop and board products are scheduled for release in July, while the server product is set for general availability in December.
2014 Next-Gen WAN SurveyWhile 68% say demand for WAN bandwidth will increase, just 15% are in the process of bringing new services or more capacity online now. For 26%, cost is the problem. Enter vendors from Aryaka to Cisco to Pertino, all looking to use cloud to transform how IT delivers wide-area connectivity.
The UC Infrastructure TrapWorries about subpar networks tanking unified communications programs could be valid: Thirty-one percent of respondents have rolled capabilities out to less than 10% of users vs. 21% delivering UC to 76% or more. Is low uptake a result of strained infrastructures delivering poor performance?